The Agent-Human Operating System

I had a catch up with Philippe Chambadal of wand.ai recently (disclaimer: I am an investor, so take my objectivity in this with a pinch of salt). I was very interested in machine learning at the time (still am), and the original pitch from wand seemed to be a high percentage of ML projects at larger enterprises were failing, and that something far more efficient, easier and more transparent was needed. I think I understood this as an enterprise-capable, no-code environment for machine learning implementation (good job I am not in marketing...wait! damn...)

Fast forward a year or two (will AC/PC ever catch on for pre-ChatGPT and post-ChatGPT?) and I was kind of expecting their focus to have been knocked to one side, or at least drowned out by the frenzy of LLMs going mainstream. I was pleasantly surprised that rather than having to pivot to gain relevance and attention, what they have done seems to be natural enough evolution and expansion of the original premise, now focused on the enterprise-capable management of humans and AI agents applied to any task or process.

The bit that stood out for me is how hybrid really means hybrid. It’s not just “humans oversee, bots execute.” Wand have built things so agents can actually delegate back to humans when a workflow needs human judgment, context, or just plain common sense. That feels obvious once you hear it - of course there are points in a process where no LLM or rules engine can reliably decide - but very few platforms bake that two-way handshake in.

It’s a clever twist. In practice, it means you don’t end up with brittle, over-automated flows and processes that break the moment something ambiguous pops up. Instead, the agent says: “I’ve taken this as far as I can - hey Brian, can you approve this exception / call this client / check this figure?” It’s not about replacing people, it’s about structuring how humans and agents pass the baton in a way enterprises can trust and audit.

And again, talking of breaking things, not every agent here needs to be a hallucinogenic LLM. Some are deterministic, some are narrow ML, and some are just smart routing engines. But when any of them hits a fuzzy edge case, they can escalate to a human within the same operating model. That’s the difference between a toy demo and something you can actually run a business on.

There are plenty of agent orchestration frameworks around, especially given we are in the (marketing?) year of the "AI Agent", but I haven't seen anyone else take the same enterprise human-agent hybrid approach that wand has. So maybe its a case of a AI framework that knows when to ask a human - and shows more humility than many of us sometimes demonstrate at work.